ComfyUI Node
Nougan MiniMax H3 KSampler π¬
MiniMax H3 sampler: SamplerCustomAdvanced-shaped (noise / guider / sigmas / latent in), CFG-free like H3 itself, with live preview and VAE + audio-VAE decode built in. model/positive are only needed to auto-build a BasicGuider when no guider is wired.
Nougan MiniMax H3 KSampler π¬
- latent_image
- noise
- guider
- sampler
- sigmas
- model
- positive
- vae
- audio_vae
- latent
- denoised
- image
- audio
βsampler_nameeulerβΊ
βschedulerbetaβΊ
βsteps8βΊ
βsteps_to_run-1βΊ
βdenoise1.00βΊ
βseed0βΊ
βdecodetrueβΊ
βtiled_decodefalseβΊ
βpreview_fps12βΊ
βpreviewtrueβΊ
βpreview_max_res1024βΊ
βpreview_frames1βΊ
βpreview_jpeg_quality80βΊ
βpreview_tiny_vaenoneβΊ
CategoryNougan/sampling
Inputs (23)
| Name | Type | Default | Description |
|---|---|---|---|
| latent_image | LATENT | β | |
| sampler_name | COMBO | euler | Built-in ComfyUI samplers (same list as KSamplerSelect). |
| scheduler | COMBO | beta | 'none' = the schedule comes from the wired sigmas input. |
| steps | INT | 81β10000 | H3 is distilled: the reference workflow uses 8 steps. |
| steps_to_run | INT | -1-1β10000 | Chainsampling: run only the FIRST N steps of the schedule, then stop partially denoised (a second sampler continues from there). -1 = all steps. Also slices a wired sigmas schedule. |
| denoise | FLOAT | 1.000β1 | β |
| seed | INT | 00β18446744073709550000 | Ignored when a NOISE input is connected. |
| decode | BOOLEAN | true | Decode the result with the connected VAE(s). Off = latent only. |
| tiled_decode | BOOLEAN | false | β |
| preview_fps | INT | 121β60 | Playback FPS of the live preview (ignored for clips with audio). |
| noiseopt | NOISE | Custom noise source (e.g. RandomNoise). Replaces the seed widget when connected (the widget hides itself). | |
| guideropt | GUIDER | Custom guider (e.g. BasicGuider). Carries the model and conditioning, so model/positive are optional with one wired. | |
| sampleropt | SAMPLER | Custom sampler (e.g. KSamplerSelect). Replaces the sampler_name widget when connected (the widget hides itself). | |
| sigmasopt | SIGMAS | Custom sigma schedule (e.g. BasicScheduler). Replaces the scheduler / steps / denoise widgets when connected (they hide themselves). | |
| modelopt | MODEL | Used to build a positive-only BasicGuider when no guider is wired (and for sigmas + the live preview). Optional when a guider is wired. | |
| positiveopt | CONDITIONING | Used with model to build the BasicGuider. Optional when a guider is wired. | |
| vaeopt | VAE | Video VAE: decodes the result into the IMAGE output and can drive the preview. | |
| audio_vaeopt | VAE | Audio VAE: decodes the audio half of H3's AV latent into the AUDIO output and drives the spectrogram preview. | |
| previewopt | BOOLEAN | true | β |
| preview_max_resopt | INT | 10240β8192 | β |
| preview_framesopt | INT | 11β1024 | Frames per step for the animated video preview (1 = single frame). |
| preview_jpeg_qualityopt | INT | 8030β100 | β |
| preview_tiny_vaeopt | COMBO | none | Tiny VAE decoder (models/vae_approx) for true-RGB previews, like KJ's Model Preview Override. Needs the decoder code: vendored tiny_vae.py or a KJNodes install. |
Outputs (4)
| Name | Type | Description |
|---|---|---|
| latent | LATENT | β |
| denoised | LATENT | β |
| image | IMAGE | β |
| audio | AUDIO | β |